Four Independence Invariants for Judgment-Quality Measurement in Agent Memory: A Specification

A number describing how well a system judges — how reliably it decides that retrieved evidence suffices, that two records conflict, that a memory should be consolidated — is uninterpretable unless the channel that produced the check is described. If the checking channel reuses the signal under test, agreement between them measures the reuse, not the judgment. This note states four invariants that make such a measurement checkable rather than merely reported. Channel declaration: the check channel is named and its source of independence declared by construction, along provenance, standpoint, substrate, or signal type. Co-error reporting: where a judgment is itself judged, at least two constructively decoupled channels are used and the rate at which they err together is reported next to the agreement figure. Layering and write-back: the write path preserves attributed contradictions, the read path may resolve them only behind an independence check, and a resolved result never re-enters the store unattributed. Proxy-label declaration: every label standing in for the quantity of interest is declared as an assumption with its scope of validity. Each invariant is given with the observation that would violate it, and with the shape of conformance evidence — paired assertions and mutation checks — that would support it.